pr-comments

name	pr-comments
description	Fetch PR review comments, triage against codebase, apply fixes, generate reply report.
arguments	[{"name":"pr_number","description":"Pull request number","required":true}]
allowed-tools	Bash, Read, Edit, TodoWrite
model	opus

PR Comments Review

Address PR review comments using MCRF with progressive confidence tracking.

MCRF Workflow

DECOMPOSE → Fetch comments, detect sources (Phase 1)
SOLVE → Categorize by action required (Phase 2)
VERIFY → Read code, validate claims (Phase 3)
SYNTHESIZE → Generate reply report (Phase 4)
REFLECT → Final confidence + caveats

Phase 1: Fetch & Parse (DECOMPOSE)

Detect repo owner/name and fetch all comment sources:

# Repo metadata
gh repo view --json nameWithOwner --jq '.nameWithOwner'

# PR overview
gh pr view $pr_number --json title,body,state,comments,reviews,reviewDecision

# Inline review comments
gh api repos/{owner}/{repo}/pulls/$pr_number/comments \
  --jq '.[] | {id, path, line, body, user: .user.login}'

# Review-level comments
gh api repos/{owner}/{repo}/pulls/$pr_number/reviews \
  --jq '.[] | {id, user: .user.login, state, body}'

Report: Count by source, parse confidence per comment.

CodeRabbit Bot Parsing

CodeRabbit embeds structured feedback inside review bodies using nested <details> HTML blocks:

• Actionable comments: marked with _Potential issue_ and severity badges
• Nitpick comments: inside <summary>Nitpick comments</summary> blocks, contain file path and line ranges
• Outside diff comments: inside <summary>Outside diff range comments</summary> blocks
• Prompt for AI Agents: inside <summary>Prompt for AI Agents</summary> — use these as implementation hints but always verify against code first

Human patterns: Short, conversational. "food for thought", "not a change on this PR", questions → typically observations not action requests.

Phase 2: Triage & Plan (SOLVE)

Categorize each comment:

Category	Criteria	Action
Actionable	Bugs, factual errors, misleading docs, test gaps	Fix
Food for thought	Observations, future ideas, questions	Acknowledge
Out of scope	Requires changes beyond PR purpose	Acknowledge + explain

Present triage table for user approval before proceeding:

#	Source	File	What	Why	Category	Confidence

Wait for user confirmation.

Phase 3: Verify & Fix (VERIFY)

Track with TodoWrite. For each actionable item:

1. Read referenced file
2. Assess claim confidence after reading code
3. Fix only if confidence > 0.9 (else flag for manual review)
4. Apply using Edit tool
5. Mark TodoWrite item complete

Constraints: No commits, no GitHub API posts, only working tree edits.

Phase 4: Generate Reply Report (SYNTHESIZE)

One section per comment with GitHub links:

• Inline: https://github.com/{owner}/{repo}/pull/{pr}#discussion_r{id}
• Review: https://github.com/{owner}/{repo}/pull/{pr}#pullrequestreview-{id}

Format:

### [Comment 1](link) — User: description ([file:L42](path#L42))
Confidence: 0.95

` ` `markdown
Fixed — what was done.
` ` `

---

### [Comment 2](link) — User: review nitpicks
Confidence: 0.90

` ` `markdown
All N items addressed:
- Item 1 description
- Item 2 description
` ` `

Grouping: Combine review-body nitpicks (CodeRabbit or human) under single section using review ID.

Final Summary (REFLECT)

1. Overall confidence (0.0-1.0)
2. Caveats: Low-confidence fixes, ambiguous comments
3. Manual review needed: Items ≤ 0.9 confidence